【作者】 黄赞武；

【导师】 魏学业；

【作者基本信息】 北京交通大学 ， 交通信息工程及控制， 2013， 博士

【摘要】 轨道电路故障是影响铁路运输行车效率、引发安全事故的重要诱因。因此,对轨道电路故障进行准确的诊断、及时的预测、科学的管理具有重要的研究意义。论文根据轨道电路的工作原理和实际使用情况对轨道电路的故障形成机理进行了深入分析,总结了轨道电路的故障类型及对应的故障征兆。在此基础上,采用故障预测与健康管理(Prognostics and Health Management, PHM)开放式分层体系,构建了一种轨道电路PHM体系结构,提出了总体解决方案,并研制了轨道电路特征参数采集及信息处理设备。论文还从以下几个方面进行了创新研究：(1)提出了一种基于小波变换的轨道电路故障征兆提取算法。针对传统FFT算法不能对非平稳信号进行时-频局部化分析的缺陷,论文利用小波变换的多尺度分辨技术对含有大量冲击干扰噪声的轨道电路信号进行分析,从时域和频域两方面对信号进行多层分解,将信号故障特征信息从干扰信号中分离出来。并通过仿真实验对比,从三种典型的小波基函数中,选取了db5小波基函数,验证了算法的有效性和精确度。(2)建立了一种基于模糊神经网络(Fuzzy Neural Network, FNN)的轨道电路故障诊断与预测模型。由于轨道电路系统具有非线性、无精确解析模型的特点,论文结合模糊推理系统易于知识表达的优点和神经网络的自学习能力,提出了一种基于模糊神经网络模型的故障诊断与预测方法。并利用和-积(Sum-Prod.)这种无参数模糊算子分别对FNN模型的故障可信度和规则激活度进行合成运算,通过仿真分析,验证了模型的有效性。(3)提出了一种基于带补偿度参数模糊算子的改进算法。通过对典型模糊算子聚合性能的分析研究,发现无参数模糊算子容易对输入信息造成遗漏,使诊断与预测结果出现较大偏差。为了解决这一问题,论文将广义概率和-广义概率积(Generalized Probability Sum-Generalized Probability Product, GPS-GPP)和广义加权均值(Generalized Weighted Average,GWA)算子用于FNN模型,并根据误差反传和梯度寻优方法推导出了两个模型的学习训练算法,通过对这两种带补偿度参数的算法和Sum-Prod.算法的分析对比,验证了基于带补偿度参数模糊算子的FNN模型比基于无参数模糊算子的FNN模型具有更高的预测精度,而且补偿参数越多,其预测精度越高。更多还原

【Abstract】 The faults from track circuits are key factors that can affect the efficiency of the railway transportation and cause safety accidents. Therefore, the accurate diagnosis, timely prognostics, scientific management of its faults has a significant meaning to the reliability and safety of train control system.According to the working principle and practical applications of track circuits, the fault formation mechanisms of track circuits are analysed in detail, and its fault types and corresponding symptoms are summarized. On this basis, using the open layered architecture of the prognostics and health management (PHM), an architecture of track circuits is built for PHM, the overall solution is proposed, and then a track circuit parameters acquisition and processing equipment is developed.The main innovation achievements of this dissertation are showed below:(1) An algorithm using the wavelet transformation is proposed for extracting fault symptoms of track circuits. As the traditional FFT algorithm cannot analyse non-stationary signal on time-frequency domain locally, an algorithm based on the multi-scale resolution of wavelet transformation is selected. The fault characteristic parameters of track circuits are got from signal with lot of impulse noises using the multilayer decomposition technique. As a result of comparing the simulations and the experiments, the db5wavelet base function is selected from three typical wavelet functions, the validity and accuracy of the algorithm is verified.(2)A model based on fuzzy neural network(FNN) is established for fault diagnosis and prognostics on track circuits. As the track circuit systems are non-linear without any accurate analytical model, with the combination of the advantages of fuzzy inference systems easily expressing knowledge and self-learning capabilities of neural networks, an algorithm using FFN model is proposed for fault diagnosis and prognostics. Two fuzzy operators without compensation parameters, Sum and Prod., are used for calculating the fault reliability and the rule activation degree respectively. Using the Sum_Prod. operators, the validity of this model is verified by computer simulation.(3) An improved algorithm based on fuzzy operators with compensation parameters is presented. Through the analysis of the aggregation performance of typical fuzzy operators, the Generalized Probability Sum-Generalized Probability Product (GPS-GPP) and the Generalized Weighted Average (GWA) operators are applied in FNN models to overcome the shortcomings. These shortcomings which come from fuzzy operators without parameters may cause the omission of input information and larger deviation of diagnosis and prognostics. The learning algorithms of these two models are derived according to the error Back Propagation(BP) and the gradient optimization methods. The comparison between the two algorithms with compensation parameters and the Sum-Prod.algorithm without parameters shows that the FNN fault diagnosis and prognostics model based on the fuzzy operators with compensation parameters has higher prediction accuracy, where more compensation parameters bring higher accuracy.更多还原